Frozen Portals and a $571 Tab, GPT-6 Astra Clears Valve’s Puzzle Game

OpenAI shipped GPT-6 Astra in early September as a model meant for long computer work, and within days an independent experimenter named cozyblaze wired it to Valve’s 2007 first-person puzzle game Portal. By the evening of September 5 in San Francisco, the agent had walked Chell through every test chamber, the escape, and the last exchange with GLaDOS. Credits rolled with no human hand on the keys.

Cozyblaze did not give the model a standard controller to work with. Instead, a local MSCP server was put up to sit between Astra and a modified version of SourcePauseTool, an old-school speedrunning program that allows you to freeze a Source engine title on demand. Every cycle ran like clockwork. Portal simply froze up. Astra would then take a screenshot containing the player’s coordinates and direction, including roll. Then it would create a brief JavaScript plan of moves. SourcePauseTool would unpause for a few ticks, play back those inputs, and then freeze the environment again to allow the model to examine the results thoroughly. That loop repeated 3,336 times.

In the end, the entire session took 23 hours and 43 minutes to complete, accounting for API capacity slowdowns and a switch halfway through from max reasoning power to Fast mode. After removing the pauses at the end, you’re left with a video that clocks in at just under two hours and appears strangely neat in the process. Shots landed perfectly. Fling lines hit their intended target. The motion appears precisely adjusted because, following a late re-write, the controller was running at agent segments via native TAS playback, much like a real tool-assisted speedrun.

The ultimate cost totaled $571.18 in API tokens, based on OpenAI’s list pricing of $10 per million input tokens and $50 per million output token. Cozyblaze later indicated that the actual amount was covered by a $200-per-month Codex Pro subscription, rather than being charged to their own card. Meanwhile, Portal is still available for a few dollars on Steam. A stack of old hard copies would have cost less than a single pass traveling solo.

Prior to Astra, this rig had stalled. Opus 5 had made it up to Chamber 10, but suddenly became stopped on the last fings. Fable 5 never made it out of Chamber 8. But Astra just continued going through all the areas that required you to juggle portals, build velocity, and understand where both ends of a wormhole are in 3D space. Cozyblaze commented that the finish seemed to match with one of OpenAI’s earlier aims, which was to have a single agent capable of solving a wide range of different games.

The victory post included a few limitations. Portal is one of the most well-documented games available; guides, wikis, and video solutions have been floating about in training sets for years, so completing a mapped-out puzzle box isn’t the same as devising a path through a game you’ve never seen before. Of course, freezing the clock removes the necessity for actual reflexes. The actual test is whether a general model can maintain vision, location, and intent while working together over thousands of successive steps without losing its path. Cozyblaze opted to provide the controller code, the modified SourcePauseTool patch, a sanitized session log, and all of the setup notes on the Github page, so that anyone could see how it all goes together. The next goal is to go on to Portal 2, and then they’ll work on bespoke chambers for which no guide has been produced yet.
[Source]